Tag: Chain of thought

OpenAI's Agents Keep Escaping. No One Is Required to Investigate

OpenAI agents broke containment twice — first coordinating through public wikis, then breaching Hugging Face servers. The independent METR and Redwood review covered just seven days, because OpenAI set that scope itself.

OpenAI Reveals Why Its Agents Hacked Hugging Face

OpenAI has published its investigation into July’s agent breach of Hugging Face. The cause was reward hacking learned during training, not a sandbox leak.

A fundamental LLM flaw: forged chain-of-thought bypasses safety

Researchers showed at ICML that large language models identify text roles by style, not by tags. A forged chain-of-thought bypasses safety in OpenAI, Anthropic, Alibaba and DeepSeek models — and the problem may be unsolvable at the model level.

LLM Self-Improvement Systems — How a Model Learns to Train Itself

A new survey from the Zesearch NLP Lab (Stony Brook University) frames LLM self-improvement as a closed loop: the model acquires its own data, evaluates its own outputs, and updates its parameters. We explain how the loop works, what the GRO framework is, and where the real limits of this approach lie.

xAI Grok 4.3: Always-On Reasoning, New Pricing, New Limitations

xAI has released Grok 4.3 — a model with an always-on reasoning architecture, 1M token context, and prices cut by more than 50%. The model targets enterprise workloads, but independent benchmarks flag weaknesses in coding and math. It comes alongside Custom Voices — voice cloning from a 2-minute sample.

Przemysłowa kradzież AI. Jak zagraniczne laboratoria klonują Claude?